Complexity Economics

Application and Relevance to Actuarial Work

A Report from the Agent-Based Modelling Working Party

Jon Palin
Nick Silver
Andrew Slater
Andrew D Smith



Introduction

The purpose of this paper is to introduce the concepts of agent based modelling to the
actuarial community. The paper will provide a brief introduction to the subject, will
look at examples of agent based models, and will identify where they might be used in
future and what the implications might be for actuarial practice and theory.

Agent based modelling (ABM) was motivated by the observation that there appears to
be a commonality between different systems in the world including ecosystems,
financial markets and the weather. Their development was facilitated by the increased
availability of computing power and the development of computational models
(Miller and Page, 2007).

ABM was not developed specifically for finance, but there is a growing interest in its
use in the field. We believe that the insights from ABM could have profound
consequences for the risk management of financial institutions, and could also be
fruitfully utilised by actuaries. However, in saying that ABM is currently in an early
stage of development whilst current models produce outputs which look like the real
markets they are a long way from producing reliable calibrated models with
predictive power.

Disclaimer: We have drafted this paper to be informative and provocative. There are
many views on agent based models, and the working party did not always reach
unanimous views. Therefore, the reader should not assume that all the authors fully
endorse every statement in this paper.


1. What is agent-based Modelling ?

1.1. The World as an Adaptive System

Proponents of agent based modelling views the economy (and indeed the world) as a
complex adaptive system, the structure of the markets, the interplay between agents
and time lags are the cause of much of the complexity and interesting behaviour we
see in the real world. Much of science (including economics) is reductionist, an
attempt to reduce the world to its basic elements. However, the interaction of these
elements causes behaviour often described as emergence - which cannot be
predicted by studying the elements themselves think of ant colonies, the brain and
economies (Miller and Page (2007)).

The agents of a complex system, often by following simple rules, form a system that
behaves qualitatively differently from the individual agents themselves. A good
example of this sort of system is an ant colony the colony as a whole can perform
complex tasks such as defending the colony from an aggressor by way of the
individual ants processing simple pieces of information.

1.2. Typical Attributes of Agent-based Models

ABMs attempts to capture this and typically have the following features:

1. Heterogeneous agents: there are a finite number of heterogeneous agents. These
agents are allowed to follow rules; these rules can be very simple (for example if
stock goes up buy) to highly complicated rules. Heterogeneity arises as groups of
agents are allowed different rules. This contrasts with neo-classical economic
models which tend to assume there is only 1 or an infinite number of
homogeneously rational agents.
2. Adaptation: agents are often allowed to adapt; i.e. individual agents can learn
during the running of the model, or are allowed to combine with other agents
(sexual reproduction) or evolve as the less successful die off and the more
successful agents flourish. Again, this contrasts with neo-classical economic
models where agents are the same through time.
3. Feedback loops: agent behaviour can have positive or negative feedback for
example if agents buy when a stock goes up this will cause a positive feedback
loop which would result in an asset bubble. Again, this contrasts with neo-
classical economic models which assume that systems return to equilibrium.
4. Local interactions: neo-classical economic models assume that all agents can trade
with each other. ABMS can allow for markets to have structure, for example
traders can only deal with brokers, there are time lags and possible geographic
affects (although the latter is not so important in financial markets).
5. Externalities: agents interact with their environment, which is in turn affected by
agents behaviour.

1.3. Parsimony and Realism

As can be imagined from the above discussion, the skill of ABM involves a playoff
between parsimony and realism. On the one extreme, many neo-classical financial
models reduce the workings of a market to a few easily solved equations. Whilst these
equations or soluble and rigorous and useful in many circumstances, they lose much
of the interesting features of real-world markets. On the other extreme, a 1-1 scale
map of the world is almost completely useless (although for some circumstances, e.g.
predicting weather it would be useful, provided you had a big enough computer).

The early results from ABM show that a complex emergent behaviour can result from
relatively simple rules amongst agents, increasing the complexity of the agents often
does not result in altering the systems behaviour. Many of the interesting life-like
results from ABMs, for example asset bubbles and crashes, income distributions, are
generated endogenously from the models due to the interaction of the 5 features
described above these are often assumed away of treated as exogenous by ne-
classical economic models.


2. Agent-Based Modelling A Brief History

2.1. Origins

The idea of agent-based modelling was developed as a relatively simple concept in
the late 1940s. Since it requires computation-intensive procedures, it did not become
widespread until the 1990s. The history of the agent-based model can be traced back
to the Von Neumann machine, a theoretical machine capable of reproduction. The
device von Neumann proposed would follow precisely detailed instructions to fashion
a copy of itself. The concept was then improved by von Neumann's friend Stanisaw
Ulam, also a mathematician; Ulam suggested that the machine be built on paper, as a
collection of cells on a grid. The idea intrigued von Neumann, who drew it up
creating the first of the devices later termed cellular automata.

Another improvement was introduced by the mathematician John Conway. He
constructed the well-known Game of Life. Unlike von Neumann's machine, Conway's
Game of Life operated by tremendously simple rules in a virtual world in the form of
a 2-dimensional checkerboard. However, it is still not a game in a traditional sense
there is only one player, and no concept of strategy. The game simply evolves.

The creation of agent-based models of social systems is often credited to the computer
scientist Craig Reynolds. He tried to model the reality of lively biological agents,
known as artificial life, a term coined by Christopher Langton.

2.2. Computational and Mathematical Organisation Theory

At the same time, during the 1980s, social scientists, mathematicians, operations
researchers, and a scattering of people from other disciplines developed
Computational and Mathematical Organization Theory (CMOT). This field grew as a
special interest group of The Institute of Management Sciences (TIMS) and its sister
society, the Operations Research Society of America (ORSA). Through the mid-
1990s, the field focused on such issues as designing effective teams, understanding
the communication required for organizational effectiveness, and the behaviour of
social networks. With the appearance of SWARM in the mid-1990s and RePast in
2000, as well as some custom-designed code, CMOT -- later renamed Computational
Analysis of Social and Organizational Systems (CASOS) -- incorporated more and
more agent-based modelling. Samuelson (2000) is a good brief overview of the early
history, and Samuelson (2005) and Samuelson and Macal (2006) trace the more
recent developments. Bonabeau (2002) is a good survey of the potential of agent-
based modelling as of the time that its modelling software became widely available.

Joshua M. Epstein and Robert Axtell developed the first large-scale ABM, the
Sugarscape, to simulate and explore the role of social phenomenon such as seasonal
migrations, pollution, sexual reproduction, combat, and transmission of disease and
even culture.

Agent-based models consist of dynamically interacting rule based agents. The
systems within which they interact can create real world-like complexity. These
agents are:
Intelligent and purposeful, but not so intelligent as to reach the cognitive
closure implied by game theory.
Situated in space and time. They reside in networks and in lattice-like
neighbourhoods. The location of the agents and their responsive and
purposeful behaviour are encoded in algorithmic form in computer programs.
The modelling process is best described as inductive. The modeller makes
those assumptions thought most relevant to the situation at hand and then
watches phenomena emerge from the agents' interactions. Sometimes that
result is an equilibrium. Sometimes it is an emergent pattern. Sometimes,
however, it is an unintelligible mangle.

2.3. Agent Based Models as an Extension of Traditional
Methods

In some ways, agent-based models complement traditional analytic methods. Where
analytic methods enable humans to characterize the equilibria of a system, agent-
based models allow the possibility of generating those equilibria. This generative
contribution may be the most mainstream of the potential benefits of agent-based
modelling. Agent-based models can explain the emergence of higher order patterns --
network structures of terrorist organizations and the Internet, power law distributions
in the sizes of traffic jams, wars, and stock market crashes, and social segregation that
persists despite populations of tolerant people. Agent-based models also can be used
to identify lever points, defined as moments in time in which interventions have
extreme consequences, and to distinguish among types of path dependency.

Rather than focusing on stable states, the models consider a system's robustness -- the
ways that complex systems adapt to internal and external pressures so as to maintain
their functionalities. The task of harnessing that complexity requires consideration of
the agents themselves -- their diversity, connectedness, and level of interactions.

2.4. Modern Applications

Agent-based models have been used since the mid-1990s to solve a variety of
business and technology problems. Examples of applications include supply chain
optimization and logistics, modelling of consumer behaviour, including word of
mouth, social network effects, distributed computing, workforce management, and
portfolio management. They have also been used to analyze traffic congestion. In
these and other applications, the system of interest is simulated by capturing the
behaviour of individual agents and their interconnections. Agent-based modelling
tools can be used to test how changes in individual behaviours will affect the system's
emerging overall behaviour.
2.5. Online Resources

A simple and accessible program for creating agent-based models is NetLogo.
NetLogo was originally designed for educational purposes but now numbers many
thousands of research users as well. Many colleges have used this as a tool to teach
their students about agent-based modelling. A similar program, StarLogo, has also
been released with similar functionality. Swarm was one of the first general purpose
ABM systems. Swarm, developed by the Swarm Development Group, uses the
Objective C programming language, and is recommended for C programmers with
little object-oriented programming experience. Swarm can also be implemented by
Java programmers, as can Ascape. Both MASON and Repast are widely used, and
EcoLab is suitable for C++ programmers. Cormas is another platform, focusing on
natural resources management, rural development or ecology research, based on the
SmallTalk language. All the toolkits described previously are based on serial von-
Neumann computer architectures. This limits the speed and scalability of these
systems. A recent development is the use of data-parallel algorithms on Graphics
Processing Units (GPUs) for ABM simulation. The extreme memory bandwidth
combined with the sheer number crunching power of multi-processor GPUs has
enabled simulation of millions of agents at tens of frames per second.

3. Simple Environment Models: Cellular Automata

Most agent-based models can be thought of in two parts: an environment and a
collection of agents. The environment sets ground rules but evolves only passively.
Agents, that live on the environment

Cellular automata is a field of academic study in its own right, and forms the
environment half of ABM (the other half being the agents that live on the
environment). While the environments used for illustration in this paper have a direct
physical interpretation they can have a much more abstract construct. For example in
a financial model the stocks that are traded are part of the environment.

What cellular automata demonstrate is how complexity emerges from very simple
rules. Consider a one dimensional line of cells, and each cell can be coloured black or
white. The colour of each cell in the next time period is determined by the colour of it
and its immediate two neighbours (one on each side) in the current time period
according to a rule that does not change across time. Between one time period and
the next the new colour of each cell is determined independently from each other
according to the rule then all cells change to their new colour.


3.1. The 256 Possible Evolutionary Rules

There are eight different combination of how three cells can be coloured either black
or white. Each combination can specify whether the cell should be black or white at
the next time period. So there are 256 distinct rules that are possible. Examples of
six rules are given below, referred (0 to 255) using an obvious binary notation. For
each rule the eight combinations are shown and below each the new cell colour that
results.

Rule 0


Rule 30


Rule 90


Rule 110


Rule 250


Rule 255


The time evolution, regardless of initial condition, of Rule 0 and Rule 255 should be
clear: all cells become either white or black. To investigate the other four rules we
consider an initial condition of one black cell with all other cells white. In the
diagrams below each line as you move from top to bottom represents the next time
period.


3.2. Example Rules

Rule 250 gives the simple behaviour of repetition.


Rule 90 gives a more complex nesting behaviour. While the pattern may be
reminiscent of a fractal remember that the diagram has been built up by applying the
same rule repeatedly line by line from top to bottom. It is not the result of a global
rule or equation, which is for example how the Mandelbrot set is constructed. And
the structure of Rule 90 is identical to Rule 250 (or 0 or 255) with just the detail
changed. The first indication that simple rules can create something interesting.


Rule 30 illustrates from simple becomes complex even further. The behaviour that
results is random, showing no regularity, even though its construction is deterministic.



As the last we look at Rule 100, which exhibits localised structures. Periods of
regularity punctuated by non-regularity; perhaps the basis for modelling the behaviour
of a financial market.



4. An Interactive Economic Model - SugarScape


Sugarscape is described in the book Growing Artificial Societies by Joshua Epstein
and Robert Axtell, and in the book they deploy the model to investigate a wide range
of social structures. Here we will look at just two. The first is growing a Pareto
distribution of wealth which fits that which is observed. The second looks at trading
between agents and the result that the actual price in a trade is often far from the
equilibrium price determined from supply and demand curves of textbook economics.

Sugarscape consists of a 50 cell by 50 cell grid (actually a torus, if an agent goes off
the left it reappears on the right, for example). For each point on the grid there is a
maximum capacity of sugar that can be supported, but at a point in time the capacity
may be less than the maximum. The initial distribution of sugar is such that the are
two mountains of sugar. The landscape is subsequently extended to include a
second commodity spice and there are two separate spice mountains. There is a
rule on how sugar (or spice) grows back each time period. In this respect the
landscape is a cellular automata. What makes it an ABM is the introduction of
agents: cellular automata + agents = Sugarscape. So there are three types of rule:
Environment Environment (eg sugar grows back after cutting)
Environment Agent (eg harvest sugar, move)
Agent Agent (eg metabolise sugar, metabolise spice, trade)

In the simulations we will look at all agents are identical, though do show variation in
their genetic makeup. (Other simulations, for example, differentiate between male
and female agents and there is a rule for reproduction, and models can include
different species of agent.) Each agent has a vision which determines how many
grid cells (1 to 6) ahead (north, south, east and west, no diagonal) it can see the
amount of sugar or spice. Each agent has a metabolism for sugar and spice, that is
how many units of sugar or spice it consumes each time period (1 to 4). Each agent
has a maximum age (which could be infinite) before it dies, if it has not already died
from starvation (when its supply of either sugar or spice reaches zero).

4.1. Initial State

These genetic characteristics are randomly assigned for each agent initially (and
similarly when a new agent is introduced when an existing one dies), together with a
random position on sugarscape (no two agents can occupy the same position) and a
random initial supply of sugar and spice between 5 and 25 units (there is no limit to
the size of each agents wealth of sugar/spice it can accumulate). Once a simulation
has been set up there is no further random element, except when a choice needs to be
made between two otherwise equal selections. So though not completely
deterministic a simulation is definitely more deterministic than stochastic post an
initial random set up.

A simulation consists of the initialisation and letting the environment and agents
follow the rules that govern them. Basically the sugarscape just wants to grow back
to maximum capacity. And basically the agents just want to move to where they can
get the best amount of sugar and spice, harvest all of it, trade with a neighbour if that
can further improve their situation, and metabolise. In each time period the agents
take their turn to move etc in an order that is random each time period (agents go one
at a time whereas cellular automata go all at once).

4.2. Model Evolution

With sugar and spice (and generally with two or more commodities) answering what
is the best allocation of sugar and spice requires a quantitative specification of
welfare to effect the move and/or trade rules. However, the qualitative description
given in the text still holds and describes conceptually how Sugarscape operates.
Readers are referred to Growing Artificial Societies for the detail which Epstein and
Axtell implemented for their model, which they considered the simplest that could be
constructed
1
. The detail is just that, a detail to make the model operate in practice and
is not something that its design rests upon. Nor are the results presented sensitive to
the formulation of the detail.


1
Much of the book can be viewed at
http://books.google.co.uk/books?id=xXvelSs2caQC&printsec=frontcover&dq=animation+IV-
2&source=gbs_summary_r&cad=0
The diagram below shows the initial distribution of sugar with each cell at its
maximum capacity.


The shades of yellow represent 0, 1, 2, 3 and 4 units of sugar. (The red dot in the
bottom left corner is a rather lonely agent.)

When spice is introduced the initial set up becomes as show in the next diagram, with
sugar in yellow and spice in orange, the red dots being agents.


The agents obey the following rules. The rules are presented in generic form, a
particular simulation will specify the necessary parameters. And the rules are
presented in English rather than computer code, which will be dependent on the
software being used.

Sugarscape grow back rule G

At each lattice position, sugar grows back at a rate of units per time interval
up to the capacity at that position

Agent movement rule M
Look out as far as vision permits in the four principal lattice directions and
identify the unoccupied site(s) having the most sugar
If the greatest sugar value appears on multiple sites then select the nearest one
(If it appears at multiple sites the same distance away, the first site
encountered is selected, the site search order being random)
Move to this site
Collect all the sugar at this new position

Multi-commodity agent movement rule M
Look out as far as vision permits in each of the four lattice direction
Considering only unoccupied lattice positions, find the nearest position
producing maximum welfare
Move to the new position
Collect all the resources at that location

Agent replacement rule R
[a,b]

When an agent dies it is replaced by an agent of age zero having random
genetic attributes, random position in the sugarscape, random initial
endowment, and a maximum age randomly selected in the range [a,b]

Agent trade rule T
Agent and neighbour compute their marginal rate of substitution (MRS); if
these are equal then end, else continue
The geometric mean of the two MRS is calculated this will serve as the price
p [more elaborate bargaining could be constructed]
The quantities to be exchanged are if p>1 then p units of spice for 1 unit of
sugar; if p<1 then 1/p units of sugar for 1 unit of spice
If this trade will (a) make both agents better off (increase their welfares) and
(b) not cause the agents MRS to cross over then the trade is made and return
to start, else end

With these rules we first look at simulation ({G
1
}, {M, R
[60,100]
}) applied when there
is just sugar (no spice). The sugarscape grows back at 1 unit per time period (recall
that agents harvest all sugar at a site when they move to it). Agents move as per the
single-commodity rule and die when they reach their maximum age which is set at
birth randomly between 60 and 100 time periods (or die of starvation).

Readers can run the simulation at by selecting experiment 3 at:
http://complexityworkshop.com/models/sugarscape.html
Agents swarm around the two sugar mountains, with some moving between the two.
Those agents with metabolism equal to 1 can survive without moving on the
lowland of sugarscape where the maximum capacity is 1 and regenerates each time
period.

4.3. Wealth Distributions

The chart below is taken from the simulation after about 100 time units, it is an
emergent structure, a stable macroscopic pattern, that statistical in nature. It shows
the distribution of wealth: along the horizontal axis there is the range of wealth of all
the agents split by decile; the vertical axis shows the number of agents in each decile.
(The complexity workshop website has NetLogo embedded as the ABM software;
NetLogo uses turtle as the label for agent.)



The distribution is Pareto: majority of agents have little wealth and a few have great
riches. Recall that all agents are created equally with random variation of their
genetic characteristics; no agents were favoured. It is the wealth distribution that is
observed in the real world. Its growth in a simple model like Sugarscape is a
powerful statement of what ABM can achieve.

4.4. Sugarscape with trade

The second simulation is ({G
1
}, {M, T}) in the sugarscape with both and spice. The
movement rule is modified as above and a trade rule has been added. The death by
old age rule has been dropped for simplicity as the focus is on trade as agents seek to
better their welfare. We can observe each trade (volume transacted and price) in a
time period and hence calculate averages for what actually happened at that time
period. We can also ask each agent its situation regarding supply and demand for
sugar or spice and corresponding price, eg how much sugar would it want to buy at
what price if there was unlimited supply. This gives the aggregate supply and
demand by price: the supply and demand curves of neo-classical economics with the
intersection being the equilibrium volume of trade and price.

A snapshot of the supply/demand vs price, actual and equilibrium, that results is given
below. We see the shape of the supply and demand is textbook, but remember that it
is an emergent feature, not an assumed structure. We also see that equilibrium is
never attained we have a model to explore the dynamics of non-equilibrium.


A movie showing a longer time evolution is available at
http://www.brook.edu/es/dynamics/sugarscape/animations/AnimationIV.mov

The charts below show the development of the equilibrium and actual price and
quantity for the first 150 time units. The key observations are:
The volume traded is always below equilibrium level (not surprising as agents
can only trade with immediate neighbours and not on a wider scale basis)
The average equilibrium quantity is 906.6, the average actual quantity is
212.5
The equilibrium price or quantity is not stable
Nor is the actual price or quantity
The actual price can be either above or below the equilibrium price (it is above
56% of the time).
The average prices are similar: for equilibrium it is 0.895 and for actual it is
0.911
The actual price or quantity exhibits greater volatility than equilibrium
The standard deviation for actual price is 0.0762 (8.4% of the mean) whereas
the standard deviation for equilibrium price is 0.032 (3.6% of the mean)
The standard deviation for actual quantity is 146.9 (69.2% of the mean)
whereas the standard deviation for equilibrium quantity is 89.4 (9.9% of the
mean)
The distribution of price is visibly non-Gaussian with evidence of fat tails.


Quantity: equilibrium & actual
0.0
200.0
400.0
600.0
800.0
1000.0
1200.0
1400.0
1600.0
1 11 21 31 41 51 61 71 81 91 101 111 121 131 141
Time
Q
u
a
n
t
i
t
y
Equilibrium
Actual

Price: equilibrium & actual
0.600
0.700
0.800
0.900
1.000
1.100
1.200
1 11 21 31 41 51 61 71 81 91 101 111 121 131 141
Time
P
r
i
c
e
Equilibrium
Actual


Distribution of actual price return
0
5
10
15
20
25
-40% -30% -20% -10% 0% 10% 20% 30% 40%
Return
C
o
u
n
t


The movement of agents is also more complicated. Agents can no longer swarm on a
single mountain top as agents need both sugar and spice to survive and these
mountains are not coincident. Movements of agents which resemble trade routes
emerge.

5. Agent-Based Stock Market Models


This section is largely a summary of an earlier and much more detailed paper (Palin,
2003, http://citeseer.ist.psu.edu/palin02agentbased.html) which contains a
mathematical specification of the model and further results and commentary.

5.1. Motivation

Stockmarkets and other financial markets share many stylised features. These include
a distribution of returns that is more peaked and fat-tailed than the Gaussian
distribution; periods of persistent high volatility; periods of persistent high trading
volume; and correlation between volatility and trading volume.

Traditional economic models have tended either to use a simple distribution of returns
such as the Gaussian and treat extreme events as outliers, or to construct a statistical
process which reproduces some of these features. But such an approach is purely
descriptive and offers no understanding of why these characteristic properties should
be seen and persist across so many markets.

To demonstrate a true understanding of the origins of these properties we would like
to build a model which (i) produces these features without them being hard-coded
into the model; and (ii) allows us to turn these features on and off by changing
parameters of the model.

5.2. Description of Le Barons model

Well focus on a model by Blake Le Baron. While there are a number of models with
broadly similar features, Le Barons model is attractive because it is makes uses of
many features of classical economics.

The model has just two assets: cash, with a constant rate of return; and an equity
which pays a random dividend in each monthly time-step. For the equity only the
dividend is specified (by a lognormal process); the price of the equity is an emergent
property of the model which arises through the interactions of the agents.

The model also has a large number of agents. Unlike the earlier Sugarscape model
there is no geographic distribution of agents: instead they all compete equally in the
same market. Agents have an initial wealth and must decide at at each time-step how
much wealth to consume, and how to invest their wealth. The agents make their
decisions in order to maximise lifetime utility. This framework allows the agents to
draw on standard results from lifecycle finance (due to Samuelson and Merton) in
order to make their decisions; but in order to do so the agents must take a view on the
distribution of equity returns over the next time-step.

The agents form their views on future equity returns by drawing on a pool of trading
strategies. Each agent has several hundred rules which turn historic market
information into a prediction. The rules are in the form of a neural network, which
allows great flexibility in potential rules and can encompass rules which we might
describe as value, growth, and momentum, as well as complex combinations of
these. The agents monitor the success of each of their rules over previous time-steps
and act upon the rule which has proved most successful. Different agents have
different lengths of memory over which they assess their rules. Some may only look
at the last ten time-steps while others may use hundreds of time-steps.

The market price is determined using a Walrasian auction Under this method each
agent says how many shares he would like to buy or sell at any given price, and the
auctioneer chooses the price so that total supply and demand are equal. This method
has been used in real stockmarkets in the past but is not in common use currently.

The model also contains elements of adaptation and evolution. If a rule has proved
unsuccessful over a number of time-steps then it well be replaced by another rule.
And agents which are unsuccessful in their trading will also be replaced.

5.3. Rational expectations price

A consequence of the grounding of Le Barons model in economic theory is that we
can calculate the rational expectations price. This is the share price that would hold
if all agents held the same rational views of the market. We find that the rational
expectations price is equal to a constant multiple of the dividend; ie that the dividend
yield is constant.

Pensions actuaries may recall that a decade ago it was common in a pension scheme
valuation to use an actuarial value of assets rather than the market value. The
discounted dividend method of assigning an actuarial value to equities is similar to
using the rational expectations price. The agent-based model allows us to consider
under which conditions the rational expectations price would or would not be seen in
practice.

5.4. Results from the model

Well look at a few different settings of the model. In each case the model is run for
many time periods to avoid any artefacts caused by the initial conditions.

We first look at the case where all agents have a long memory: they choose which
rule to use based on performance over a period of around twenty years (240 monthly
time-steps). In this case (Figure 1) we find that the share price does converge to the
rational-expectations price, and the share price in the chart below is lognormal (since
it just a multiple of the dividend which is itself lognormal). We also find in this case
that there is no trading, since each agent is happy to maintain the same constant mix
of cash and equity.


Figure 1: The long memory case. The rational expectations price is observed.

We can contrast this with the all memory case where agents memories vary from
six months to over twenty years. In this case (Figure 2) we see significant deviations
from the rational expectations price with bubbles and crashes. We also find that in
contrast to the long memory case there is significant trading volume as agents
change their views in response to the changing market. The trading volume is
correlated with changes in share price, with trading volume being particularly high
during the instability around time 10000 and 10200.


Figure 2: The all memory case. Significant deviation from the rational expectations
price.

We would like to know whether the deviation from the rational expectations price in
the all memory comes from the diversity of agents memory lengths or just from the
presence of short memory agents. To address this we consider the short memory
case where all agents have memories that range from six months to three years. This
shows rapid cycles of bubbles and crashes.


Figure 3: The short memory case. Very rapid cycles of bubbles and crashes.

5.5. Cautions and criticisms

Le Barons model is a partial success. It allows us to investigate how a share price can
emerge from the actions of individual agents, and it shows how changing a parameter
of the model (agents memory length) can move from lognormal share prices to
stylised features of markets with fat tails, bubbles and crashes, and persistent
volatility.

But while we can get a good qualitative result it is difficult to find a calibration which
gives a good quantitative calibration. Changing memory length gives a sudden shift
from rational expectations price to a wild caricature of the real world (the kurtosis of
returns is an order of magnitude too high).

Also different models give different apparent reasons for features of real markets. One
model claims that a calibration with fast evolution of agents and strategies is
necessary to see fat tails; another shows fat tails without any evolution.

Although agent-based stockmarket models as they currently stand do not seem ready
as an investment tool they remain a very useful tool for investigation. The processes
which drive stock returns are not well understood with much commentary,
particularly in the popular media, being glib and applied only after the event. The
complex dynamics seen in LeBarons model are valuable in showing that simple
explanations of stock prices in terms of a single event are probably wrong.

6. Agent Based Models and Finance

When the authors of this paper were setting out on our careers, actuarial valuation
methods were typically deterministic, based on expected outcomes minus implicit or
explicit margins for prudence. On the pensions side, this is the funding approach to
liability valuation (Exley et al (1997)), while in insurance liabilities were valued by
reference to reliable yields (a calculation still required). The underlying
methodology was challenged in the 1990s by a series of papers arguing in favour of
market valuations
2
. These were backed up on strong modern financial economic
foundations, a discipline which hitherto had had little impact on actuarial practice.
These papers were highly influential in convincing the profession to alter practice, for
example the adoption of market based valuations, which are now generally common
practice in actuarial valuations and for accounting purposes in pensions and insurance.
Alongside this move to market valuation was the growth of computing power which
allowed the deterministic approach to be replaced with a stochastic one.

Agent-based modelling is not a mere tweak to neo-classical finance theory, but an
alternative world view which matches many participants intuition of what real
markets look like not necessarily a guarantee of their veracity. For example, agnet
based models are inherently lumpy, while much of classical finance makes
assumptions about infinitely divisible agents and investment markets in order to apply
differntial calculus. Some agent based models consider market and intrinsic values
separately, and explain how these two numbers diverge for significant periods of time.
The agent-based model generates bubbles, crashes and high volatility endogenously,
rather than exogenously as in the efficient market hypothesis (Farmer (2001)). For
anyone who works with real markets, they are intuitively appealing.

6.1. Financial Uses of ABM

The use of agent based models so far has been dominated by aspects of explanation
rather than prediction. Consideration of explanation and prediction need to
differentiate between qualitative and quantitative, issues of calibration being more
important for the latter. For example a model (agent based or otherwise) of financial

2
For example Exley et al (1997) and Sheldon and Smith (2004)
markets needs to generate bubbles as they have been observed. A model which does
not is clearly defective in either explanative or predictive capability. Having
generates bubble behaviour it is for calibration techniques (which may not currently
exist) to achieve a realist scaling between model and observation. Models which have
passed the qualitative test may fail the subsequent quantitative test.

6.2. Modigliani, Miller and Agents

A series of papers 1990s laid the foundations of what we now call market consistet
modelling. These papers applied the emerging discipline of finance economics to
actuarial theory. For example, a seminal paper was Exley et al (1997) which applies
the Modigliani Miller theorem (that the value of two firms is the same irrespective
of their financial structures (Modigliani Miller (1958))) to defined benefit pension
scheme valuation and investment strategy.

The paper argues that a shareholder of a company with a defined benefit pension
scheme has three equivalent ways of changing his asset mix:

1. Altering his directly held assets
2. Changing the balance sheet of the company
3. To modify pension fund asset strategy (assuming the shareholder has power to do
this)

The implications of this equivalence are profound as they imply that it is impossible
to achieve an optimal investment portfolio, as the shareholder can alter his own
investments instead of the pension funds.

The paper is now going to look at how we might take an agent based approach to
modelling a pension fund. The aim of this exercise is not to describe a realistic model,
but to question whether ABM could give rise to a different outlook to Modigliani-
Miller.

Doyne Farmer of the Sante Fe Institute constructs two kinds of traders, a seasonal
trader and a technical trader. There is no movement in the underlying value of the
commodity in the market. The seasonal traders buy and sell in a predictable pattern
they could represent, for example, farmers or even companies who need to meet
emissions targets. The technical traders are purely in it for the money and are allowed
to develop a variety of trading strategies; the more successful a strategy is, the more
capital the trader gets and the more they influence the market. The result is shown in
Figure 4. What happens is that the market becomes efficient after about 5,000
iterations when the technical traders make money off the seasonal traders and iron
out the predictable seasonal fluctuations. But after that, as you see, the model
suddenly goes mad. This is because the technical traders, who have now acquired
practically all the capital (as shown in graph B), start trading against each other,
devising ever more sophisticated trading strategies which work for a while until
another trader develops a better strategy (Farmer (2001)).

6.3. Pension fund as agent?

To answer this question let us imagine a scenario in which an equity and bond
portfolio both gave the same performance over a 5 year period, but equities increased
by 50% over the 1
st
3 years and then declined to meet the level of the gilts after 5
years i.e. one not unlike our current experience.

At the end of this period, the shareholder is equally well off if he invested directly in
bonds or equities. However, the outcome of shareholder value if the pension fund
invested in equity or gilts might be quite different. If the pension fund invested in
equity, at some point there would be an actuarial valuation, which would show a
surplus this could give rise to a reduced contribution rate and possible change to a
more aggressive investment approach. The reduction in contributions would mean
that the company has more cash available to invest in other activities. The positive
balance sheet may boost the companies share price, making capital cheaper which it
then might invest favourably. The subsequent drop in equity value (exacerbated by
the more aggressive investment policy) may put the scheme into deficit, causing the

Figure 4 Behaviour of market with 2 types of trader (Farmer 2001)

company to close the scheme, maybe boosting share value but also alienating the
workforce.

This is the potential action of the pensions fund we could in theory build up
predictable investment rules for the pension fund, which could then be modelled as an
agent. However, many pension funds tend to face the same issues at the same time. If
we model pension funds as agents we could introduce them into the a Farmer or
LeBaron type model, which are likely to challenge the Modigliani-Miller
equivalence.

This scenario above is oversimplified the aim is to demonstrates that the pension
schemes decision to invest in equity or gilts could have an effect on the pension
scheme, the value of the company and shareholder value quite different to his decision
to change his personal investments.

7. Implications for Actuarial Practice

7.1. Early Stages

Agent-Based Modelling is still at an early stage. What we have now is a collection of
example computer models. The inputs are a collection of agents with assumed
behaviour patterns, and ways of interacting. The outputs include simulated paths of
financial variables, which in most cases share prices.

The models we have, and future we may build, allow us to conduct many
experiments. With a combination of carefully controlled experiments and the growing
mathematical insights from the field of dynamical systems, we can aspire to a better
understanding of what in the model is most important at driving the outcomes. For
example, is the absolute number of agents critical, or is there some point beyond
which the model is large and adding extra agents does little to change behaviour.
How could we measure the diversity of behaviour among agents? Is it the absolute
level of irrationality or the difference between rationality of different agents that
matters most? To what extent do different forms of irrational behaviour cancel out in
observed market behaviour? Does the behaviour of all agents contribute equally to the
modelled behaviour, or does a cadre of leaders emerge whose behaviour is more
influential on the market as a whole?

7.2. The Need for Calibrations

We are far from understanding these issues at the current state of knowledge. But this
is likely to improve as experimentation and analysis proceeds. An ability to rank
inputs in order of importance is a pre-requisite for empirical calibration. The ranking
tells us what aspects of real trader behaviour should be captured in order to build a
useful model of the economy. Having determined the most important inputs,
calibration might proceed by surveys, interviews and trading on market simulators.
All this provides a way to calibrate the behavioural aspect of the models.

Without a calibration, the model tells us nothing concrete about the real world.
Models of efficient markets have existed for some time. ABM provides us, in
addition, with many alternative models of different ways markets could be inefficient.
Building an abstract computer model of inefficient markets cannot prove that markets
are inefficient. Efficiency, or otherwise, is an empirical question for which disciplined
observation of the real world is critical. That is why calibration is so important. A
fully calibrated and tested ABM, with demonstrable predictive power, can add to our
knowledge about real markets, not just to our knowledge of computer models. We are
not there yet, so a view that ABM will one day provide insight into market efficiency
remains, for now, a hunch, albeit a widely held hunch among ABM enthusiasts.

7.3. Rigour in testing Goodness of Fit

Informal model tests involve checking that the model accounts for observed data
patterns. A more rigorous process takes account of the number of parameters
estimated. The argument is that, with a large enough parameter count, a model can
reproduce any data set. The trick in classical statistics is to find a parsimonious model
than explains as much as possible of the observed data with a small number of
parameters. Neo-classical optimising agents fit well into this framework; a
parsimonious objective function is specified, and all aspects of behaviour are in
single-minded pursuit of the chosen objective. Specifying the behaviour of a non-
optimising agent requires an exhaustive list of responses to any eventuality. Without
any over-arching optimisation at work, the potential number of rules is vast. As a
result, ABM can generate many thousands or millions of parameters. In a statistical
sense, such high parameter counts imply a high threshold for the models ability to
explain real world outcomes.

7.4. Forecasting Ability

Let us now conduct a thought experiment, and imagine we have a calibrated and
tested ABM. What could we do with it?

The most obvious use is market prediction. A rejection of market efficiency is a
rejection of the notion that returns are unpredictable. So an ABM that consistently
predicts market prices (more accurately than a random walk) would pose a substantial
challenge to theories of efficient markets. Orthodoxy might fight back, for example by
rationalising observed biases as a reward for risk, but if the rewards are large enough
then a latent risk is a less plausible explanation. The implications for portfolio
construction and investment strategy are profound. We do not know how many of
these are in use already; the owner of such money-making machine may be reluctant
to publicise the fact. Conversely, if ABMs do not ultimately give better forecasts
than a random walk, then the ABM has instead illustrated how irrational agents can
nevertheless contribute to market pricing a result that would also be of considerable
interest albeit less lucrative.

It is less clear how other financial theories would need adapting. The Black-Scholes
model for option pricing does not rely on market efficiency. It does rely on being able
to trade continuously, without transaction costs, and on having accurate volatility
forecasts. These conditions do not fully hold in real markets, yet the Black-Scholes
and related models are still widely used. In our view, it is unlikely that the increasing
use of ABM would dent the use of Black-Scholes model, but it could well provide
new insights into the elements driving an appropriate choice of volatility parameter.

7.5. Adapting Modigliani and Miller

The work of Modigliani and Miller on capital structure has had a profound impact on
recent actuarial thought, particularly in pensions. The underlying arguments are based
on arbitrage and do not require markets to be efficient. However, to exploit the
arbitrage of a violation of Modigliani and Miller, would require an arbitrageur to take
offsetting long ands short positions in two shares with equal core businesses but
different capital structures or different pension plan investments. Such trades are in
practice impossible to execute, because companies do not come in such convenient
pairings. We can instead turn to equilibrium arguments to imply the same (M&M)
conclusions, but now we have introduced the additional hypothesis that economic
agents are trying to optimise something. Take away optimising agents, and the
arguments become much more delicate. Everything becomes very model dependent;
for example under Modiglinai and Miller (and excluding tax, bankruptcy or benefit
leaage) an extra 1000 pension fund contribution should have no effect on a
companys share price. The usual lesson taken from this argument is that optimal
contributions much therefore take account of tax, bankruptcy and leakage effects.
However, in an agent based model, we cannot easily unpick the effect of contributions
on share prices. ABM might suggest we need to know a great deal more in depth
about the investor community in order to recommend pension contribution rates.

7.6. Feedback Loops

Some ideas of market inefficiency are already embedded in regulatory thinking. If
markets are efficient, and prices reflect best information, then there can be little
market benefit from regulatory interference in trading volumes. If a regulator prevents
an institution from selling a risky asset, then the share price of the institution itself
falls, so the fall in aggregate value of the economys assets is unchanged by the
regulation. One could even argue that a regime discouraging fire sale of assets expose
policyholders or pension plan members to further risk, because in a default event it is
policyholders or plan members who have to scrape around the remaining assets to
recover some of their promised benefits.

Under agent based models, however, positive feedback loops may operate to
exacerbate the effect of crashes. Relaxation in solvency regulation, and corresponding
easing of the pressure on institutions to sell risky assets, could mitigate the market fall
itself. This increases aggregate economic wealth and reduces some of the refinancing
costs that otherwise might have applied. As calibration technology develops, we may
well see a day when ABM provides meaningful support to regulators searching for the
best way to protect plan members or policyholders in turbulent financial times.


References

Beinhocker, E (2007) The origin of wealth Business Books

Bonabeau, E (2002) Agent-based modelling: Methods and techniques for simulating
human systems Icosystems Corporation

Brealy, R and Myers, S (2007) Principles of corporate finance McGraw-Hill
Education
Copeland, T and Weston, F (2004) Financial theory and corporate policy (4th
Edition) Addison-Wesley

Exley, C Mehta, S and Smith A (1997) The financial theory of defined benefit pension
schemes British Actuarial Journal 3, IV 835-966

Fama, E and French, K (1992) The cross section of expected stock returns The journal
of finance Vol XLVII, No 2.

Farmer, D (2001) Toward agent-based models for investment Association for
Investment Management and Research

Holland, J (1998) Emergence Oxford University Press

Mandelbbrot, B and Hudson R (2006) The (mis)behaviour of markets Profile Books

Mandelbrot, B (1963) The variation of certain speculative prices The Journal of
Business, 36(4):394419

Mandelbrot, B (1967) The variation of some other speculative prices The Journal of
Business, 40(4):393413

Miller, J and Page, S (2007) Complex adaptive systems Princeton Paperbacks

Modigliani, F and Miller M (1958) The cost of capital, corporation finance and the
theory of investment American Economic Review, 48, 261-297

Palin, J (2002) Agent-based stockmarket models: calibration issues and applications
University of Sussex

Ralfe, J Speed, C and Palin, J (2003) Pension and capital structure: why hold equities
in the pension fund? Society of Actuaries

Sharpe, W Alexander, G and Bailey, J (1995) Investments Prentice Hall Fifth Edition

Sheldon, T and Smith, A (2004) Market consistent valuation of life assurance
business Institute of Actuaries

Taleb, N (2007) The black swan Random House

Varian, H (1999) Intermediate microeconomics Norton 5th Edition